Victor Luo

Led scaling law and feature interaction research for a full-scale replacement of the main ranking model with a new SOTA model, collaborating with three senior researchers to achieve a 5% increase in Watch Live Duration (AWLD) org-wide.

Experimented with SOTA model architectures, including transformers and stacked networks, to improve scaling performance by 5x parameters and boost UAUC by 2%.

Spearheaded the redesign of the entire retrieval system, developing a novel generative retriever model that leverages Trinity clustering to learn explicit user interest, enhancing diversity metrics by +10% and AWLD by 1.4%.

Designed and implemented Transformer-based sequence modeling in the first-stage ranking model, improving UAUC by 0.6% and increasing watch duration by 1% through 100 user interactions.

Integrated a SOTA lifelong memory module to extend sequence modeling contextual range, resulting in a 2% improvement in watch duration.

Developed a new LLM feature to improve user understanding on device detail pages by implementing a billion-scale embedding retrieval system, streamed via Kafka and Kubernetes, to achieve improved latency and serve 7M daily visits.

Developed and scaled large data pipelines with a scalable backend leveraging AWS S3 and API Gateway, automating merchandiser actions on Amazon Device web pages that reach over 360M customers annually.

Generated $20M quarterly revenue by automating merchandiser actions on Amazon Device web pages, directly contributing to business growth and efficiency.

Developed and deployed the first-ever multi-agent LLM framework and application in Tensorflow, enabling LLM agent cooperation and improving cross-entropy by 50% through agent fine-tuning with QLoRA.

Led a research partnership with Google, applying directed-cyclic graph simulations for medical information alignment to advance SOTA LLM capabilities.

Contributed to a conference paper on text generation fairness models, aligning experiments tracking with training parameters to ensure robust and unbiased AI outcomes.

Trained and fine-tuned a PETER transformer with Pytorch, leveraging distributed GPUs and multi-task learning to provide personalized explanations on terabytes of review data.

Outperformed the GPT-3 baseline in accuracy (BLEU) and text quality for personalized explanations, demonstrating superior model performance and efficiency.

Co-founded BOUND, an ML platform recognized as a "Top 30" solution, securing initial funding to optimize company meetings with employee feedback.

Led product development with a 4-person engineering team, overseeing the entire lifecycle from concept to deployment.

Grew the platform to serve 35 B2B clients, demonstrating strong market penetration and business development capabilities.